On my projects at Target, I've had a good experience using DuplicateRecordFields coupled with a library like overloaded-records or generic-lens. This lets us define records that share field names without needing to use prefixes or multiple modules, while still being able to access and modify those fields easily thanks to the OverloadedLabels extension.
With this approach, you get a lens for every single record field without name conflicts. If you have a bunch of records with a field called id, you'll be able to access them using #id as a lens without running into name conflicts with other records *or* with normal identifiers (like the id function). You can then access fields with `record ^. #fieldName`, set them with the .~ operator and compose them with other lenses and traversals.
When we first out together a system like this I was worried that it would break type inference and cause weird error messages, but it had not caused *any* problems in practice. Haskell's records went from being a royal pain to being at least as convenient as any other language I've used, and sometimes more (eg when combined with traversals).